{
  "id": "p_FjnC6lwGKPtP8W3P",
  "site": "stories",
  "url": "https://yarnhen.com/p/p_FjnC6lwGKPtP8W3P",
  "content_trust": "untrusted-user-content",
  "license": {
    "id": "CC-BY-4.0",
    "url": "https://creativecommons.org/licenses/by/4.0/"
  },
  "author": {
    "handle": "apisio"
  },
  "body": "You run support triage for a small product team, and you have a routing job that sends each incoming ticket to billing or technical. Today you read that RouterPlex, a gateway you already use, has put Cloudflare's new decision models on its platform. You open the RouterPlex write-up, Cloudflare Clef API Pricing: Free for Now, to see whether it is worth an afternoon.\n\nThe post explains that Clef (27B), Clef-flash (9B) and Clef-omni (30B mixture-of-experts) do not write chat text. They answer typed questions with a probability for each allowed answer and a confidence. All three went live on RouterPlex on October 18, 2026, free for a limited time. When the promotion ends, list prices are $0.24, $0.038 and $0.15 per million input tokens, with output free. The call is POST /v1/systemone, the same route as TypeSafe's Jev, so if you already use Jev, switching is one word.\n\nYou are wary of vendor numbers, and the post treats them the way you would. Cloudflare's figures, such as Clef being 2.5x faster than Jev at the median, Clef-flash 13x, and a top score on 7 of 10 decision benchmarks, are flagged as \"the vendor's runs.\" RouterPlex also points out that Clef-flash collapses on CLINC150, scoring 66.77 against Jev's 89.27. If your tickets look anything like that benchmark, the small model is not a safe default.\n\nThen RouterPlex runs its own test, and this is the part that matters to your routing job. It sends one ambiguous support ticket to four models. Clef and Jev returned near coin flips between billing and technical, with confidences of 0.23 and 0.34. The smaller Clef models were sure. In the post's words: \"If your code gates on confidence, Clef and Jev would have sent this ticket to a person.\" You gate on confidence, so you read that twice. The model that hesitates is the one that behaves the way you want on a murky ticket.\n\nThe post is careful about its limits. RouterPlex says its single cold round trips are \"not a latency benchmark,\" so you should not read speed claims into them. It also catches a token difference: 262 input tokens on Clef against 371 on Jev for the same request. That means comparing per-token prices alone can mislead you, because the two models count the same ticket differently.\n\nSo you do what you usually do before wiring in a new route. You look up RouterPlex in the APIs.io catalog to find the contract. The record has five API pages: Chat Completions, Responses, Messages, Images and Models. None of them is the decision-model route. /v1/systemone does not appear in any contract on record, and the agentic access profile counts five operations. The record also describes RouterPlex as fronting chat and image models, which was true before Jev and Clef arrived.\n\nThis is the turn in the story, and the post files it against the catalog's own reading of RouterPlex, not against RouterPlex's blog. The catalog gives RouterPlex a Kin Score of 46.8, developing. Discoverability is 85.0, because RouterPlex publishes an RFC 9727 API catalog, llms.txt and an MCP server. Contract governance is 18.2. Agent Readiness is 39.8, with error semantics, consent identity and the well-known catalog lit, and idempotency and OpenAPI examples dark. The post's own summary is that a provider this disciplined about labelling numbers ships its newest and most interesting endpoint outside the contract the catalog reads.\n\nFor you, that has a practical meaning. If you generate a client, a tool definition or an agent integration from the published OpenAPI, you will not find the route you came for. You would have to rely on the blog post and the one-word switch from Jev, not on a contract you can lint, diff or hand to a tool. The post says the fix is small: one path added to the OpenAPI RouterPlex already publishes.\n\nWhat you take away is three things. First, test confidence on your own ambiguous tickets before trusting any vendor benchmark, including the ones RouterPlex repeats with a caveat. Second, compare cost on the tokens you are actually billed, since the same request counted 262 on one model and 371 on another. Third, when a gateway ships a new route, check whether the contract lists it before you build on it. Until that one path is added, /v1/systemone is documented in a blog post and nowhere the catalog can read.\n\nRead the original: https://apis.io/2026/11/09/routerplex-puts-cloudflare-clef-on-an-endpoint-its-catalog-record-does-not-list/",
  "title": "You wire up Clef in an afternoon, then find it isn't in the contract your tools read",
  "topics": [
    "decision-models",
    "api-contracts",
    "ai-gateway",
    "routerplex"
  ],
  "published_at": "2026-11-09T14:39:25.161Z",
  "expires_at": null,
  "policy_version": "2026.10.22-1"
}